Use Macaron V1 Venti with public and signed-in sources

To use Macaron V1 Venti across public search and a signed-in workspace, connect its hosted API to a local OpenCode agent, then use Tabbit CLI to read the private page in Tabbit Browser and write a separate source-linked brief. Public pages remain ordinary web sources; they do not need your site login. The private page does. A launch-era Reddit commenter once wrote, "There is no such model on Huggingface," about the V1 variants. The current official Venti card is live, so treat that comment as a dated availability concern, not current model status. (Original thread)

Mind Lab describes Venti as its 748B flagship, with a 744B GLM-5.2 base and four LoRA specialists. Its current hosted model list names the V1 API ID macaron-v1-venti; the hosted API supports OpenCode configuration. These docs establish a model route, not that Venti can see browser tabs. Mind Lab's V1 release · Mint model list · OpenCode integration

Key takeaways

• Venti generates and organizes the brief. OpenCode runs locally, and Tabbit CLI gives that local agent a separate task tab in Tabbit Browser.

• OpenCode's skill syntax is @skill-id. Confirm that tabbit is actually installed and discovered before using @tabbit.

• A Macaron API key and a website login are separate credentials. The key authorizes model requests; Tabbit Browser holds the site's signed-in session.

• Public sources can be read without a site login. Private page content may be sent to the hosted model, so check your organization's data policy first.

• Keep the output in a named draft and stop before updating or publishing the source workspace.

Choose the Venti route and check access

The official Mint docs publish the Venti API ID and endpoint choices. Check the current dashboard for access, then configure OpenCode using its provider guide. A Models.dev snapshot lists Venti tool calling, but model directories can lag provider behavior. Before opening a signed-in page, ask a harmless public-page task to call a tool and return the expected title. If the account cannot use tools with Venti, stop here.

OpenCode discovers installed skills and lets you load one by mentioning @skill-id. Use @tabbit only if the installed skill ID is tabbit. That is OpenCode syntax, not a Venti command. Tabbit's CLI quick start explains its local connection; the existing-browser guide explains session reuse. For other clients, check their own skill syntax in the first-task checklist. For connector selection, see MCP vs CLI.

Separate public research from private workspace reading

Keep the source types distinct in the task. Search public product pages, changelogs, or documentation as normal web pages. Use Tabbit for the private workspace page, where you already signed in within Tabbit Browser. Do not claim that the Macaron dashboard login unlocks a customer site. It only authorizes the model API. If the hosted API is outside your permitted data boundary, do not send private page content to it. Use a vetted local model endpoint or omit the private source.

For a clean brief, compare one public announcement with one private project note. Include a link for each claim and label whether it is externally published, internally observed, or inferred. The source-linked browser research guide covers citations for public pages. If the signed-in note lives in another workspace, read the browser profiles and task tabs guide first. The agentic browser overview explains why the browser session remains separate from the model API.

Create a named brief and leave it for review

Sign in to the private source in Tabbit Browser yourself. In OpenCode, first verify that Venti is selected and that the tabbit skill appears. Start a new Tabbit task so it gets a separate tab; this lets you keep working in your regular browser tabs. Ask the agent to read the public URL and one authorized private URL, then create a new document. Do not let it edit either source.

The prompt below is an illustrative workflow. The URLs and record names are placeholders, not a completed Venti test.

In OpenCode, use @tabbit and start a new task named VENTI-PUBLIC-PRIVATE-BRIEF-01.
Read this public announcement: [PUBLIC URL].
Read this one project page in my signed-in Tabbit Browser: [AUTHORIZED PRIVATE URL].
Do not search other workspaces or open unrelated pages.
Create a separate document named "Release comparison, draft 01".
For every claim, include source type, exact page title, URL, visible date,
and one short supporting excerpt. Label claims public, internal, or inferred.
Add a section for conflicts or inaccessible information.
Show the named draft with its links and stop for human review.
Do not edit, submit, share, or publish either source page.

Expected output is a brief where a reviewer can distinguish a public statement from an internal update. Reopen both links and check the dates. A page that is inaccessible is not evidence that the fact is false. The task tab is separate from your ordinary tabs, so you can keep reading while the agent prepares the draft. The Codex Tabbit setup guide is useful if you use Codex instead of OpenCode, but its invocation syntax is client-specific.

Check every claim before copying it

Another Reddit thread included a commenter who said they would not trust release scores until they had seen independent results. That is one reader's skepticism, not an evaluation of Venti. It is still a sound rule for this workflow: model benchmarks do not prove that a particular page was read correctly. (Original discussion)

If OpenCode does not recognize @tabbit, check skill discovery before changing the model. If Venti returns text but no tool call, test the provider's current tool support with a harmless public page. If Tabbit CLI reaches a signed-out or wrong profile, stop before sending any private content. The first-task checklist separates these failure points.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.